Dual Attention Transformer for Identifying Alzheimer’s Disease in MRI Scanning Applications
Abstract
Alzheimer's disease (AD) is a slowly progressive neurodegenerative disease that is very hard to spot in its early stages because of the small changes in the shape of brain structures. MRI is a standard non-invasive imaging method for diagnosing Alzheimer's disease (AD), but traditional deep learning models, especially CNNs, often fail to accurately classify MRI scans because they do not account for both global contextual dependencies and local fine-grained features. We will solve this problem by proposing a Dual Attention Transformer (DAT) architecture that combines spatial and channel attention with a Vision Transformer (ViT) core. Channel attention concentrates on modal differences, whereas spatial attention assists you in focusing on those anatomical regions of body that are significant to disease. The model was trained and tested using the ADNI MRI dataset which consists of groups of people with Alzheimer;s disease (AD), Mild impaired cognition (MCI) and people whose brain is cognitively normal (CN). Our approach performed better than baseline CNNs and standalone ViTs with better accuracy, precision, and F1 scores. Ablation experiments were an additional source of information which indicated that each attentional sub region has a different role to play. Finally, DAT model can be recommended as an appropriate choice to diagnose AD at the early stage since it is convenient and easily comprehensible. It might be applied in clinical pipelines of MRI in the real world.